Direction-of-Arrival (DOA) estimation is a fundamental task in acoustic signal processing and is used in source separation, localization, tracking, environment mapping, speech enhancement and dereverberation. In applications such as hearing aids, robot audition, teleconferencing and meeting diarization, the presence of multiple simultaneously active sources often occurs. Therefore DOA estimation which is robust to Multi-Source (MS) scenarios is of particular importance.

In the past decade, interest in Spherical Microphone Arrays (SMAs) has been rapidly grown due to its ability to analyse the sound field with equal resolution in all directions. Such symmetry makes SMAs suitable for applications in robot audition where potential variety of heights and positions of the talkers are expected. Acoustic signal processing for SMAs is often formulated in the Spherical Harmonic Domain (SHD) which describes the sound field in a form that is independent of the geometry of the SMA. DOA estimation methods for the real-world scenarios address one or more performance degrading factors such as noise, reverberation, multi-source activity or tackled problems such as source counting or reducing computational complexity.

This thesis addresses various problems in MS DOA estimation for speech sources each of which focuses on one or more performance degrading factor(s). Firstly a narrowband DOA estimator is proposed utilizing high order spatial information in two computationally efficient ways. Secondly, an autonomous source counting technique is proposed which uses density-based clustering in an evolutionary framework. Thirdly, a confidence metric for validity of Single Source (SS) assumption in a Time-Frequency (TF) bin is proposed. It is based on MS assumption in a short time interval where the number and the TF bin of active sources are adaptively estimated. Finally two analytical narrowband MS DOA estimators are proposed based on MS assumption in a TF bin.

The proposed methods are evaluated using simulations and real recordings. Each proposed technique outperforms comparative baseline methods and performs at least as accurately as the state-of-the-art.